PreprintPDF Available

Multi-Agent Organizational Structure: specification and validation

Authors:
Preprints and early-stage research may not have been peer reviewed yet.

Abstract and Figures

To design complex Multi-Agent Systems (MAS), the use of high-level abstraction concepts such as roles, protocols, and groups, makes the task relatively easier and closer to the reality of the system or domain that we want to simulate or create. These concepts were introduced through the multi-agent Organizational Models (OM) which can be seen at two levels: i) an abstract level which is the Organizational Structure (OS) and ii) a concrete level which is the Concrete Organization (CO) to be deployed and executed. This article is dedicated to the formal specification and validation of the identified roles in the Organizational Structure before a real deployment to generate systems with organizations exhibiting good qualities. Also, a proposal of a set of steps to automatically generate the code of the designed agent will be presented. In addition, the evaluation of some organizational aspects will be done at the abstract level makes it possible to adjust the design to remedy any failures before a real instantiation. The application domain used in this context, is a Cooperative Information Gathering System (CIGS) for travel organization.
Content may be subject to copyright.
Page 1/21
Multi-Agent Organizational Structure: specication
and validation
Issam Bouslimi
University of Jendouba
Research Article
Keywords: Multi-Agent Systems, Organizational Model, Organizational Structure, Cooperative Information
Gathering System
Posted Date: April 5th, 2024
DOI: https://doi.org/10.21203/rs.3.rs-4197325/v1
License: This work is licensed under a Creative Commons Attribution 4.0 International License. 
Read Full License
Additional Declarations: No competing interests reported.
Page 2/21
Abstract
To design complex Multi-Agent Systems (MAS), the use of high-level abstraction concepts such as roles,
protocols, and groups, makes the task relatively easier and closer to the reality of the system or domain
that we want to simulate or create. These concepts were introduced through the multi-agent
Organizational Models (OM) which can be seen at two levels: i) an abstract level which is the
Organizational Structure (OS) and ii) a concrete level which is the Concrete Organization (CO) to be
deployed and executed. This article is dedicated to the formal specication and validation of the
identied roles in the Organizational Structure before a real deployment to generate systems with
organizations exhibiting good qualities. Also, a proposal of a set of steps to automatically generate the
code of the designed agent will be presented. In addition, the evaluation of some organizational aspects
will be done at the abstract level makes it possible to adjust the design to remedy any failures before a
real instantiation. The application domain used in this context, is a Cooperative Information Gathering
System (CIGS) for travel organization.
1 Introduction
MAS design can be considered according to two approaches: an agent-centered approach (ACA) or an
organization-centered approach (OCA) [12]. The ACA approach is essentially concerned with the denition
of the internal structure of the agents in terms of mental states (Logic, Believe, Desire, Intention, Reactive
…). It focuses on the micro level of the system to be modeled and does not master the interaction
between the systems components. The system behavior is supposed to emerge, and the designer has
very few controls over it. The OCA, however, tries to overcome these shortcomings by two means: i) by
focusing on the observable characteristics of the agents in terms of capacities instead of their internal
structure; ii) by introducing organizational abstractions that make it possible to structure and regulate
agent interactions for better organization of tasks, limitation of conicts and reduction of communication
cost [13]. For this purpose, concepts such as role, group, norms, laws, or interaction protocols are
proposed [1]. Accordingly, the organizational approach can be viewed as a technique that governs
individual behavior of agents to make them converge towards the global objectives to be achieved.
Individual behaviors must comply with norms and rules xed by the organization design [5].
Considering all these advantages, it becomes clear that following an organizational approach (instead of
an agent-centered one) facilitates the design of complex, open and secured MAS [10]. In addition, all
these facts explain the emergence of organizational models in MAS, among which we mention AGR [13],
MOISE+[14], GAIA2[20] and an OM proposed in [2]. Also, a framework called MASQ was proposed to
design Organization Centered Multi-Agent Systems (OCMAS) [11]. In these models, we distinguish two
abstraction levels: The Organizational Structure (OS), and the Concrete Organization (CO).
In this paper, we will focus on the OS: how we can design the internal behavior of the roles? Is there a way
to automatically generate the code of the role class? and how we can evaluate and validate this level
before moving to the instantiation?
Page 3/21
Using a Cooperative Information Gathering System (CIGS) [3] through a case study of travel organization,
we will proceed as follow: i) identication of the roles involved in the system and their specication using
the Object Petri Nets (OPN), ii) simulation of the internal behavior of these roles using the Renew tool.
This simulation is used to check the functional criteria like liveness and termination of the process, iii)
propose a set of steps to automatically generate the code of the designed role, iv) quantitative validation
of organizational aspects such as exibility, eciency, and robustness [4, 16] using metrics that can be
calculated a priori.
The rest of the paper is structured in the following way. In section 2, we will present the state of the art
related to the formal specications of roles and the validation of the OS. In Section 3, we will introduce
the used case study: the Cooperative Information Gathering System (CIGS). Section 4 gives an overview
of the OPN and the design of identied roles using this formalism. Section 5 focuses on the simulation of
these roles using the Renew tool and the validation of behavioral criteria. The steps to link the design to
the implementation are proposed in Section 6. Quantitative validation of the Organizational Structure is
presented in section 7. Lastly in section 8 conclusions are made.
2 Related work and positioning
Between 2000 and 2010, signicant attention was devoted to the organizational approach adopted for
designing and developing MAS. Several Organizational Models have been proposed, and readers are
referred to [27] for a detailed list. A survey of relevant works dealing with formal specication of multi-
agent organizations is presented in [30].
These studies have proposed formalisms to formally specify the organizational dimension. Among these
works, we mention rst [16] since the evaluation criteria of the Organizational Structure proposed in this
manuscript in section 6 are those presented in this work. The authors, inspired by [25, 26], propose a
formalization of the OS and its properties using oriented graphs. This formalization allows for a thorough
evaluation of the robustness, exibility, and eciency of the OS before actual deployment, and enables
the selection of the conguration that best ts the scenario to be implemented. However, this
formalization suffers from two shortcomings:
Only the macro level of the organization is addressed. The proposed method does not refer to the
internal behavior of each role. The OS is than presented as a set of roles and their mutual
interactions without any formalization of the actions carried out by each role.
Interactions between roles are limited to the types of relationships that exist between them: authority,
coordination, or control. Consequently, during implementation, there is no mention of the objects
exchanged between these roles or the protocol governing these exchanges.
In [28], an OM was proposed as part of a comprehensive approach called MACODO. Organizational
abstractions are formalized using the Z language. This language provides programmers with support to
develop dynamic organizations that adapt to environmental changes. However, this specication is
Page 4/21
closely tied to the application domain of the MACODO approach, which is trac monitoring. Additionally,
no macro view of organizational structuring has been proposed.
One additional study worth mentioning is [29] as it is closely related to our proposal of modelling multi-
agent organizations with Petri Nets. The authors use PN to model a system they describe as simple. This
formalization has only enabled the verication of a single criterion, which is the system is deadlock free.
In addition, transition inputs consist of simple tokens, whereas in reality, the entities being manipulated
are objects. Moreover, a detailed description of the internal actions of the agents has not been addressed.
Although the organizational approach focuses on high-level abstract concepts without referring to
implementation details, it is interesting to specify the functioning mode of roles (and consequently
agents in a subsequent stage).
In this study, we will demonstrate that using OPN for role specication offers the following advantages
compared to other formalisms:
The formal specication of the micro level of the Organizational Structure is ensured through the
formalism. For each role, a detailed description of all its actions and interventions is provided.
Formalizing the micro level of the Organizational Structure also dictates the mode of exchange
between roles by enforcing a specic protocol. These protocols are implicitly described through the
interactions of each network/role with other networks/roles. Therefore, the language provides a
specication that encompasses both the micro and macro levels of the Organizational Structure.
PN allow the verication of functional criteria before real system deployment. The criteria to be
veried by simulation in section 5 are termination, nal state accessibility, and liveness.
Since OPN combine object-oriented programming concepts with PN, and since all internal role
actions and interventions/interactions are modelled, an automatic transition from design to
implementation is possible. Although what will be proposed in section 6 is not very comprehensive or
elaborate, feasibility remains attainable. Indeed, combining a OPN editor with a set of transition rules
enables the generation of the code skeleton related to the modelled role.
Finally, the design of a Cooperative Information Gathering System to verify organizational criteria is
presented. Although the criteria come from [4, 16], but the aim is to propose a comprehensive
approach for specifying and validating the abstract level of the multi-agent organizations.
3 Overview of the Organizational Model for Cooperative Information
Gathering
The case study used in this work is a Cooperative Information Gathering System (CIGS). A CIGS is an
organization of agents that work together to collect and share information for a common goal. These
agents are programmed to collaborate and communicate with each other to eciently gather and
process data from various sources. The system starts with an initial user query, which will be
decomposed into elementary tasks. These tasks will be distributed among agents, allowing for parallel
Page 5/21
processing and faster information retrieval. By working together, these agents can cover a wider range of
sources and gather more comprehensive data than a single agent working alone.
The CIGS was designed by following a multi-level Organizational Model. The problem treated by this
system is the trip organization, where the user will introduce a list of parameters and the system will
provide a result based on what is possible and user preference. For more details, we refer readers to [6, 7].
To specify and describe the functioning of the CIGS and more precisely how agents interact, we choose to
model the system with an organizational perspective. Introducing organizational abstractions like roles
and protocols makes it possible to structure and regulate agents’ interactions to organize their
functioning, to limit conicts and reduce communication cost [9]. Using these two abstractions, allows
the design of the organization without worrying about the micro level composed of agents, but rather
focusing on the macro level composed of roles and their interaction protocols. At run time, an agent is
then likely to play several roles, and a role can be played by several agents.
Thanks to a functional analysis, we have identied the roles which necessarily intervene in any CIGS:
- The Mediator decomposes and/or reformulates the initial user query, supervises the execution of each
elementary task and constructs the result.
- The Coordinators are in charge of elementary tasks. They coordinate with the matchmaker to nd
translators able to retrieve information. After the reception of that information, they communicate it to the
mediator and to other coordinators if needed.
- The Matchmaker provides references towards external informational agents able to carry out elementary
tasks.
- The Translators (or wrappers) are external agents found by the matchmaker and in charge of retrieving
information. They act as an Application Programming Interface (API) to allow the interaction between an
information source and the coordinator.
4 Role representation with Object Petri Nets
4.1 Overview of Object Petri Nets Formalism
In order to represent the identied roles in our CIGS, we had chosen the Object Petri Nets (OPN)
formalism. This choice is motivated by the fact that OPN is a parallel systems specication language
allowing us to formalize, in one hand, the concurrent activities of the roles intervening in a CIG process,
and in the other hand, the roles’ actions, its interventions, the invariants, the coordination rules, its
resources and the access authorizations to each resource.
Object Petri Nets (OPN) [18] are a formalism coherently combining Petri Nets (PN) technology and the
Object-Oriented (OO) approach. While PN are very suitable for expressing the dynamic behavior of a
Page 6/21
system, the OO approach permits the modeling and the structuring of its active (actor) and passive
(information) entities. In a conventional PN, tokens are atomic, whereas in an OPN, they are presented as
objects. As any PN, an OPN is made up of places, arcs, and transitions, but in an OPN, they are labeled
with inscriptions referring to the handled objects. More precisely, an OPN features the following additional
characteristics:
- Places are typed. The type of a place is a (list of) type of an (list of) object(s). A token is a value
matching the type of a place such as a (list of) constant (e.g. 2 or ‘hello’), an instance of an object class,
or a reference towards such an instance. The value of a place is the set of tokens it contains.
- Arcs are labeled with parameters. Each arc is labeled with a (list of) variable of the same type, as is the
place that the arc is connected to. The variables on the arcs surrounding a transition serve as formal
parameters of that transition and dene the ow of tokens from input to output places. Arcs from places
to a transition determine the possible condition of the transition: a transition may occur (or is possible) if
there exists a binding of its input variables with tokens lying in its input places.
Each transition is a complex structure made up of three components: a precondition, an action and
emission rules. A transition may be guarded by a precondition, i.e. a side-effect free Boolean expression
involving input variables. In this case, the transition is only permitted by a binding if this binding
evaluates the precondition to be true. Passing a transition through depends on the precondition, on the
location of tokens and on their value. Most transitions also include an action, which consists in a piece of
code in which transitions’ variables may appear and object methods be invoked. This action is executed
at each occurrence of the transition, and it processes the values of tokens. Finally, a transition may
include a set of emission rules i.e. side-effect free Boolean expressions that determine the output arcs
that are activated after the execution of the action.
4.2 Role specication with OPN
We use the six following principles (explained with reference to the mediator in Fig. 2) to represent the
behavior of a role by an OPN:
- The rectangles delineate the roles with which it interacts. Only communication places appear, and which
are called interface. This interface is an input place for one role and an output place for the other.
- action associated with the transition represents the internal actions carried out by a role,
- the interventions (I1, I2,...) which are transitions from transmitting role to a receiver role, are represented
by arcs labeled by KQML performatives [19].
- a condition that holds true for each accessible marking beginning with an initial marking is known as an
invariant. There exist numerous Petri Nets analysis techniques to deduce them.
- the grey places are the resources,
Page 7/21
- the authorizations are represented by arcs, which connect transitions to resources.
The internal behavior of the Mediator role is show in Fig. 2. It interacts with the roles User and
Coordinator. This interaction is represented by the six interface places on the top of the gure.
At the left side we found the Resources used by the Mediator which are the Informational Model (IM) and
the Task Model (TM). These two models belong to a whole modeling process explained with details in [7].
The
QuUsMe
place (for
QuestionUserMediator
) represents the initial state of the mediator. It’s expressed
by an
Ask
performative. The nal state is shown by the
AnMeUs
place, and it expresses the reception of
the nal result via a
tell
performative.
Three distinct phases, discernible within the network, comprise the Mediator's behavior:
The user's query can be accepted by the Mediator, who can then reformulate it, break it down into smaller
questions, and send them to the coordinators or the user. The left portion of the net is composed of
transitions
T1
and
T2
. Following receipt of the user's question, the Mediator reformulates it (Transition
T1
) considering the user's preferences as stated in the instant message as well as the question's context.
The Transition
T2
is broken down by the Mediator into smaller questions that are sent to the Coordinators
and maybe the user when the query has been reformulated. We utilize a dotted arc to indicate that
participation in this interaction is optional. Then, the Mediator waits for the answers to these sub-
questions (
AwaitingResults
place).
The responses to the sub-questions can be gathered by the Mediator thanks to the middle portion of the
net, which is made up of transitions
T3
and
T4
. Upon receiving a result from a Coordinator or a user, the
Mediator records it (using
T3
or
T4
) in the
AllTheResults
place and noties the
ResultsNotication
place
of the existence of new results.
The portion of the network on the right, consisting of transitions
T5
and
T6
, examines the collection of
results that have been saved (
T5
). If those results are deemed adequate, they are combined (
T6
) to
produce the result, which is then presented to the user. Should the outcomes be insucient, the Mediator
continues to wait for additional results. As soon as fresh data are received, it will begin an analysis (
T5
)
once more. It is notied of this event if one or more tokens are present in the
ResultsNotication
location.
T1.T2.((T3|T4).T5)
n
.T6
is the language of the net, which provides all conceivable behaviors, where n is
the total number of answers to the questions the coordinators and user submitted that the mediator
received. The nal state is reached by the execution of
T6
transition which consumes
WaitingResults
token.
5 OPN simulation
In addition to the formal specication of the OS roles given above, we performed a simulation to verify
the behavioral aspect of the mediator role previously specied. We have chosen as a simulator the Renew
Page 8/21
tool (the Reference Net Workshop) developed at the Department of Computer Science at the University of
Hamburg. Renew is freeware (http://www.renew.de), written in Java. It allows the graphic editing of the
PN which is also used as a synoptic during the simulation.
The simulation of the behavior of the Mediator role reveals two possible scenarios but which both lead to
the desired end state (expressed by a token in place of the OPN of the previous section and which
informs the user of the presence of a nal response to his request). These scenarios, presented in Figs. 3,
correspond to the participation or not of the user when formulating the nal response. The simulation
allowed us to validate that the network was not blocked in both cases. It should be noted that the T5
transition has been split into two sub-transitions (T5.1 and T5.2) since the Renew tool does not allow us
to express the validation conditions.
To verify that all transitions can be crossed, we ran a step-by-step simulation for both scenarios. This
simulation allowed us to validate the near liveliness of the network. Figure4 shows this stepwise
execution of the model relating to the second scenario (with user interaction T4).
The simulation made it possible to verify the following properties:
- Unnished termination: termination is not guaranteed because we have an analysis loop of the response
(((T3|T4).T5)n) which is intentionally modeled as an iteration which allows it to take into account
responses as they arrive.
- The accessibility of the nal state: there is an evolution of the network leading from the submission of
the question (initial state) to the resolution of this question (nal state).
- Liveness of our network: there is for each action (transition) a conguration of the network allowing its
execution.
6 Role implementation
Once the termination, nal state accessibility, and liveness properties have been validated through
simulation, it would be interesting to proceed to the implementation stage and generate code that
complies with the specication.
Specifying roles using OPN allows the designer to generate the skeleton of the class related to that role.
Although the process is currently manual, automation is feasible by creating an API between the E-
NetObject [21] tool, which is a OPN editor, and a MAS development platform such as JADE [22],
AgentBuilder [23], or MADKIT [24], enabling automatic generation of role skeletons. The API will start by
parsing the XML le generated by the E-NetObject tool, which represents a snapshot of the edited net.
This step will identify all designed entities such as transitions, manipulated objects, and interventions.
The second step will involve applying a set of transitioning rules that will translate the identied entities
in the rst step into Java code of the class relevant to that role.
Page 9/21
The choice of the Java language is arbitrary since most software editors for MAS are based on this
language. Each MAS development platform has a set of predened functions for each agent, so the API
must by default integrate these methods. One of these methods is the one we will call Life (the naming
varies from one platform to another), which implements the agent's behavior throughout its lifecycle.
The transitioning rules from an OPN to Java code can take the following form:
- Each transition is a method that takes the incoming object as a parameter and produces the outgoing
object as a result.
- Each interaction is a message sent or received with a parameter which is the KQML performative.
- Each object manipulated by the role, and which is external to the system to be implemented is a
parameter of the Life method. For example, in our CIG context, external objects are resources and the
initial request formulated by the user.
- The Life method will implement the networking language, which is in our case:
T1.T2.((T3|T4).T5)
n
.T6
as described in section 3. This method takes as parameters all external objects to the system. In our case,
these objects are the Info Model, the task Model (as Ontology type in our case), and the user's initial
question (String type).
By applying these transition rules, the skeleton of the resulting Java class is shown in the following Fig.
5.
7 Quantitative assessment of the Organizational Structure
The two organizational levels of the OM need to be accompanied with tools and metrics, to make the
evaluation process possible, before a real deployment of the multi-agent system. In this section, we will
focus on the quantitative evaluation of the rst organization level which is the Organizational Structure.
We refer readers to [8] for qualitative evaluation of the communication in the concrete organizational
level.
7.1 Evaluation criteria
The criteria used to evaluate the performance of the OS are inspired from [4, 16]. The authors applied
quantitative concepts from graph theory in order to assess the structure used to implement the MAS. We
will briey introduce these criteria but for a complete explanation we refer readers to [4, 16, 9].
The organizational properties which are evaluated are:
- Robustness: how the OS is stable in an unpredictable environment
- Flexibility: the capability of an organization to adapt to changing circumstances
Page 10/21
- Eciency: how to achieve the global goal of the system with the minimum of resources.
To evaluate these properties, the authors introduced three equations to measure specic graph-theoretical
aspects of organizational structures which are:
- Connectedness: expresses how strongly the roles are linked with one another
- Economy: how we can keep a system connected with the minimum of links
- Univocity: expresses the absence of redundant links within the same role
Each one of these structural aspects is calculated within one of these three structural dimensions which
described the nature of links between each role:
- Authority dimension: it implies that a role can delegate a task to another role.
- Coordination dimension: which represent the knowledge exchange between roles.
- Control dimension: the fact that a role must monitor the activities of another role.
The organizational properties, which are robustness, exibility, and eciency, are measured from graphs
representative of the different roles and the different types of links that exist between them. The authors
provided a vector of values which represent the optimal results that we can obtain for each one of these
properties. During the two next sections, we will present the role graphs from where we will make the
calculations, and we will compare our results with these optimal vectors of values.
7.2 Role graphs
In this section, we will introduce the representative graphs of the different roles of the Organizational
Structure according to the three structural dimensions of control, authority, and coordination. Since the
CIGS was conceived to resolve the problem of travel organization, the OS was strictly dependent on the
problem resolving process. As shown in Fig. 6, the problem is decomposed into four elementary tasks
which are the visits, the accommodation, the transport, and the weather report. The organization of
accommodations is extremely related to the transport as they should have the same arrival and departure
dates. Each elementary task will search information from a different Information Sources. As a result, the
OS includes the following roles:
- A Mediator, overseeing the CIG process,
- As many Coordinators as there are basic tasks in the trip organization problem. In our case there are
four: visits, the accommodation, the transport and the weather report. These coordinators will be titled, in
the role graph, respectively, Coordinator 1, Coordinator 2, Coordinator 3 and Coordinator 4.
- Since a Matchmaker can be specialized in more than one eld, we supposed that we will have three.
Therefore, we have three Matchmakers respectively rated Matchmaker1, Matchmaker 2 and Matchmaker
Page 11/21
3.
- As many Translators as there are elementary tasks. In our case, we chose to associate a translator with
each coordinator. We will then have four translators labeled respectively Translator 1, Translator 2,
Translator 3 and Translator 4.
To facilitate the reading of the representative diagram of the different types of links between the roles
identied in our OS, we have chosen to present a graph for each structural dimension (control, authority,
and coordination) already introduced in section 5.1:
The "control" dimension (see Fig. 7): the mediator has a control link over all the coordinators allowing him
to assess the provided results.
The “authority” dimension (cf. Figure 8): The Mediator has both a link of authority and control with the
coordinator roles, which allows it to both delegate the sub-objectives to them and assess the returned
results. In turn, the Coordinators have a link of authority on the one hand over the Matchmakers by
requiring them to provide the addresses of information sources likely to contain the desired result, and on
the other hand over the Translators by ordering them to query and return partial results from these
sources.
The "coordination" dimension (see Fig. 9): The Coordinator can interact with another Coordinator in the
case of solving interrelated problems which are expressed through a coordination link. Message feedback
as responses to requests between the different roles is presented by a coordination link. Finally, the
Translators publish their skills to the Matchmakers, which also translate into a coordination link.
7.3 Results and interpretations
The results obtained from the application of the equations provided in [4, 16] as well as the maximum
values which optimize the three structural properties are given by the following tables:
Table 1
Robustness
STRUCTURAL PROPERTY OPTIMAL VALUES OBTAINED VALUES
Economycoordination 0 0,94
Univocityauthority 0 0,91
Connectivitycoordination 1 1
Page 12/21
Table 2
Flexibility
STRUCTURAL PROPERTY OPTIMAL VALUES OBTAINED VALUES
Economycoordination 0 0,94
Connectivityauthority 0 1
Connectivitycoordination 1 1
Table 3
Eciency
STRUCTURAL PROPERTY OPTIMAL VALUES OBTAINED VALUES
Economycoordination 1 0,94
Economycontrol 1 1
Economyauthority 1 1
From these results, and by comparing them with the vector of values that optimize the three
organizational properties, we can draw the following conclusions:
- The organization is not strong. In fact, two out of three values are far from the optimal values, which are
Economy-coordination and Univocity-authority. Let us recall in this context that robustness induces the
system's ability to resist in the face of unpredictable events such as the failure of an agent, playing a
given role, to provide a certain service. However, and as we have dened the current OS, very little
redundancy is present, so there is no alternative for delegation in the case of failure. During execution, a
Coordinator may judge that the result returned by a Translator is irrelevant and consider the recruitment
of another Translator: which is not considered in our modeling. The lack of robustness is also due to the
specialization of our agents which are therefore not interchangeable. In this context, and to x this
weakness, we must integrate the possibility of an auto-adaptation of agent goals in case of unpredictable
events as described in [17].
- The organization is not very exible because we have a very centralized, very economical organization
and therefore with very little redundancy in terms of coordination. The values of the Economy-
Coordination and Connectivity-Authority dimensions are far from the optimal values. In our case, and to
restrict the information exchange between agents, and to reduce the cost of communication, the
recruitment of agents is centralized at the level of the Mediator and Coordinators who exercise authority
over the rest of the agents. In short, exibility has been sacriced for the benet of the economy in terms
of interactions. A solution to the lack of exibility in the MAS organization acting in uncertain
environments is proposed in [15]. This can be the subject of future works.
Page 13/21
- The organization is very ecient. Indeed, the three values of this criterion are almost optimal. This is
explained by the fact that we have optimal modeling at the level of each dimension, so a minimum
structure to coordinate, control and direct. Only the mediator and the coordinators have authority over the
other agents, and the coordination and control links are reduced to the essential. The effectiveness of the
organization as dened in [3] is on its optimal value as the goals are well accomplished by the agents.
8 Conclusion and future work
In this paper we have proposed a formal specication of the identied roles in the abstract level of the
Organizational Model (the Organizational Structure) of a Multi-Agent System. The chosen formalism was
the Object Petri Nets. This formalism enabled us to formally present the internal actions, interventions,
and the rules of sequence of the various roles. This representation formalism also has the advantage of
being simulated in order to validate the properties of the model. We were then able to validate the
behavioral aspects of the roles via a simulation with the Renew simulator. Properties such as non-
blocking and liveliness of the model have been veried. We also proposed a way to automatically
generate the code of each role based on its OPN specication. We are conscient that the automatic
transitioning from the design to the implementation require more technical studies, but we are convinced
that is feasible. In addition, we made a quantitative evaluation of our OS based on graph-theoretical
measures. This evaluation concern organizational properties such as exibility, robustness and eciency
and was carried through a Cooperative Information Gathering System. The objective was to provide a
comprehensive approach for dening and verifying the abstract level of the multi-agent organizations,
even though the evaluation criteria originate from [4, 16].
Declarations
Author Contribution
Only I.B. wrote the whole article
References
1. A. Artikis and J. Pitt, 2001, A Formal Model of Open Agent Societies.
In: 5th International Conference
on Autonomous Agents (AA’01)
, Montreal-Canada, pp. 192-193.
2. Afsaneh Fatemi, Kamran Zamanifar, and Naser Nematbakhsh, 2001. Adaptive Team-Based Multi-
Agent Organizational Model: A Case in Rescue Systems.
International Journal of Computer Science
and Information Technology
3.2, pp. 165–175.
3. Daniel D. Corkill, Edmund H. Durfee, Victor R. Lesser, Huzaifa Zafar, and Chongjie Zhang,2011.
Organizationally Adept Agents. In
Working Notes of the AAMAS-11 Workshop on Coordination,
Organizations, Institutions, and Norms in Multiagent Systems
(COIN).
Page 14/21
4. Grossi, D., Dignum, F., Dignum, V., Dastani, M., Royakkers, L., 2006. Structural evaluation of agent
organizations.
5th International Joint Conference on Autonomous agents and Multiagent systems
,
pp. 1110-1112.
5. Hosny Ahmed Abbas, Samir Ibrahim Shaheen, Mohammed Hussein Amin, 2015. Organization of
Multi-Agent Systems: An Overview.
Journal of Intelligent Information Systems
, Vol. 4, No. 3, pp. 46-
57.
. I. Bouslimi, Khaled Ghédira and Chihab Hanachi, 2006. An Agent-based Organizational Model for
Cooperative Information Gathering,
In: Proceeding of the second International Conference on Signal-
Image Technology and Internet-based Systems, SITIS
, Hammamet, Tunisia, pp. 414-425.
7. I. Bouslimi, Chihab Hanachi, Hassan Tout and Khaled Ghédira, 2008. Coordination framework for
Cooperative Information Gathering.
In: International Journal of Advanced Intelligence Paradigms,
Inderscience Publishers
, Vol. 1, No. 1, pp. 60-79.
. I. Bouslimi, C. Hanachi and K. Ghédira, 2014. An Experimental Evaluation of Communication in an
Organization-Based Multi-agent System.
IEEE/WIC/ACM International Joint Conferences on Web
Intelligence (WI) and Intelligent Agent Technologies (IAT)
, pp. 72-78.
9. Ines Thabet, Issam Bouslimi, Chihab Hanachi and Khaled Ghedira, 2011. A Multi-agent
Organizational Model for Grid Scheduling.
In: KES International Conference (KES-AMSTA
2011
),Manchester, UK, Vol. 6682, pp. 148-158.
10. Jensen, Andreas & Villadsen, Jørgen, 2013. A comparison of organization-centered and agent-
centered multi-agent systems.
Articial Intelligence Research. 2. 10.5430/air.v2n3p5
9.
11. J. Ferber, T. Stratulat, J. Tranier,2009. Towards an integral approach of organizations in multi-agent
systems: the MASQ approach. Chapter book of Multi-agent Systems: Semantics and Dynamics of
Organizational Models in Virginia Dignum (Ed), IGI.
12. J. Ferber and O. Gutknecht, 1998, A Meta-Model for the Analysis and Design of Organizations in
Multi-Agent Systems.
In: 3rd International Conference on Multi-Agents Systems, (ICMAS’98), IEEE
Computer Society
. Paris, France.
13. Jacques Ferber, Olivier Gutknecht, and Fabien Michel, 2004. From Agents to Organizations: an
Organizational View of Multi-Agent Systems.
In: Agent-Oriented Software Engineering (AOSE) IV
, P.
Giorgini, Jörg Müller, James Odell, eds, Melbourne, (2003), LNCS 2935, pp. 214-230.
14. J.F. Hübner, J. S. Sichman, O. Boissier, 2002. MOISE+: towards a structural, functional, and deontic
model for MAS organization.
In: the First International Joint Conference on Autonomous Agents &
Multi-Agent Systems, (AAMAS’02)
, Bologna-Italy pp. 501-502.
15. Keogh, Kathleen, and Liz Sonenberg. 2020. Designing Multi-Agent System Organizations for Flexible
Runtime Behavior.
Applied Sciences
10, no. 15: 5335. https://doi.org/10.3390/app10155335
1. Loris Penserini, David Grossi, Frank Dignum, Virginia Dignum, Huib Aldewereld. 2009. Evaluating
Organizational Congurations.
In Proceedings of The 2009 IEEE/WIC/ACM International Conferences
on Intelligent Agent Technology (IAT-09), IEEE CS Press
, Milano, Italy.
Page 15/21
17. Matteo Baldoni, Cristina Baroglio, Roberto Micalizio, Stefano Tedeschi, 2023. Accountability in multi-
agent organizations: from conceptual design to agent programming. In Autonomous Agents and
Multi-Agent Systems 37. https://doi.org/10.1007/s10458-022-09590-6
1. Sibertin-Blanc, 1985. High-level Petri nets with Data structure.
6th European workshop on Petri nets
and applications
, Espoo (Finland).
19. Tim Finin, Richard Fritzson, Don McKay, and Robin McEntire, 1994. KQML as an agent
communication language.
In Proceedings of the third international conference on Information and
knowledge management (CIKM '94)
. New York, NY, USA, pp. 456–463.
20. Zambonelli, N. R. Jennings and M. Wooldrige, 2003. Developing Multi-Agent Systems: The Gaia
Methodology.
In ACM Transactions on Software Engineering Methodology
. pp. 317-370.
21. Frédéric Raclot, David Andreu, Thérèse Libourel Rouge, Robin Passama. E-NetObject: Un Editeur de
Réseaux de Petri à Objets. 02180, 2002, pp.40. irmm-00269413f
22. https://jade.tilab.com/, 10/10/2023.
23. https://www.agentbuilder.com, 12/01/2024.
24. https://www.madkit.net/madkit/index.php, 01/10/2023.
25. D. Grossi, F. Dignum, M. Dastani, and L. Royakkers. Foundations of organizational structures in
multiagent systems. In AAMAS ’05: Proceedings of the fourth international joint conference on
Autonomous agents and multiagent systems, pages 690–697, New York, NY, USA, 2005. ACM.
2. D. Grossi, F. Dignum, V. Dignum, M. Dastani, and L. Royakkers. Structural aspects of the evaluation of
agent organizations. In COIN@ECAI 2006, 2006.
27. Dignum, V. (Ed.), 2009. Handbook of Research on Multi-Agent Systems: Semantics and Dynamics of
Organizational Models: Semantics and Dynamics of Organizational Models. IGI Global.
2. Weyns, D., Haesevoets, R., & Helleboogh, A, 2010. The MACODO organization model for context-
driven dynamic agent organizations. ACM Transactions on Autonomous and Adaptive Systems
(TAAS), 5(4), 16.
29. Celaya, Jose & Desrochers, Alan & Graves, Robert. (2007). Modeling and Analysis of Multi-agent
Systems using Petri Nets. Journal of Computers - JCP. 4. 1439 - 1444.
10.1109/ICSMC.2007.4413960.
30. S. Sabeg, T. M. Maarouk and M. El Habib Souidi, "Formal Specication and Verication for
Organization-based systems : A Survey,"
2022 4th International Conference on Pattern Analysis and
Intelligent Systems (PAIS)
, Oum El Bouaghi, Algeria, 2022, pp. 1-6, doi:
10.1109/PAIS56586.2022.9946660.
Figures
Page 16/21
Figure 1
Formal specication of the Mediator Role with OPN
Page 17/21
Figure 2
Scenario with user interaction vs Scenario without user interaction
Figure 3
Step-by-step simulation for scenario 1
Page 18/21
Figure 4
This image is not available with this version.
Page 19/21
Figure 5
Mediator java class code
Page 20/21
Figure 6
Trip organization elementary tasks
Figure 7
Control links of the Organizational Structure
Page 21/21
Figure 8
Authority links of the Organizational Structure
Figure 9
Coordination links of the Organizational Structure
ResearchGate has not been able to resolve any citations for this publication.
Article
Full-text available
This work proposes a notion of accountability for multi-agent systems, that supports the development of robust distributed systems. Accountability is grounded on responsibility, and encompasses both a normative dimension, and a structural dimension. For realizing robust distributed systems, conceived as agent systems or organizations, it is necessary to keep a right level of situational awareness, through the introduction of the means for gathering and propagating accounts, upon which actions can be taken. This paper presents a formalization of accountability, including the accountability lifecycle, for the design of robust agent organizations. Particular attention is given to the interplay of accountability and goals, by describing typical patterns in which accountability affects the state of an agent’s goals and vice versa. We illustrate the practical aspects of the proposal by means of JaCaMo (Boissier et al. Sci Comput Program 78(6):747–761, 2013. https://doi.org/10.1016/j.scico.2011.10.004).
Article
Full-text available
We address the challenge of multi-agent system (MAS) design for organisations of agents acting in dynamic and uncertain environments where runtime flexibility is required to enable improvisation through sharing knowledge and adapting behaviour. We identify behavioural features that correspond to runtime improvisation by agents in a MAS organisation and from this analysis describe the OJAzzIC meta-model and an associated design method. We present results from simulation scenarios, varying both problem complexity and the level of organisational support provided in the design, to show that increasing design time guidance in the organisation specification can enable runtime flexibility afforded to agents and improve performance. Hence the results demonstrate the usefulness of the constructs captured in the OJAzzIC meta-model.
Article
Full-text available
In this paper, we addressed the problem of modelling complex cooperative multi-agent organizations. Proposed team-based model is inspired from the Schwaninger's model of intelligent human organizations. Considering time and agents' identity features is the main important factor in the mentioned model. In our considered problem domain, rescue systems, rapid task handling is a main requirement. Proposed model supports it through fast initial team formation, greedy capability-based coalition formation, and using the nearest neighbours' resources. Adaptation via reorganization makes the model appropriate for dynamic environments.
Chapter
Full-text available
A multi-agent system can be analyzed and specified as an organization consisting of roles and their relations. The performance of an organization depends on many factors among which the type of its organizational structure, i.e., the set of relations holding between its roles. This work focuses on the structure of organizations and addresses the issue of the analysis, evaluation, and comparison of organizational structures which can contribute to develop general methods for the assessment of multi-agent systems’ performance. Specifically, quantitative concepts from graph theory are used to provide numerical analyses of organizational structures. It is argued that these analyzes can be used for evaluating to what extent an organizational structure exhibits some characteristic properties such as robustness, flexibility and efficiency.
Chapter
In this chapter, we stress the importance of thinking a MAS in all its aspects (agents, environment, interactions, organizations, and institutions), using a more integral vision. We show that a genuine organizational approach has to take into account both the environment and the institutional part of MAS societies. Then, we propose the MASQ (Multi-Agent System based on Quadrants) meta-model, which constitutes an abstraction of the various aspects of an OCMAS (Organization Centered Multi-Agent Systems), extending AGR (Agent/Group/Role). MASQ is based on a four-quadrant framework, where the analysis and design of a system is performed along two axes: an interior/exterior dimension and an individual/collective dimension. We give a conceptual definition of this approach and we will show that it is possible to apply it to practical models.
Book
Organizational design plays a critical role in the development of larger and more complex multi-agent systems (MAS). As systems grow to include hundreds or thousands of agents, those in the field must move from an agent-centric view of coordination and control to an organization-centric one. The Handbook of Research on Multi-Agent Systems: Semantics and Dynamics of Organizational Models brings together views and approaches from different communities including MAS, computational organizational simulation, social simulation, organizational theory, and cognitive science, to present the intersections of the different topics. Containing chapters by renowned international authors, this book provides a comprehensive volume on the state-of-the-art on formal semantic models and practical applications for dynamic formalisms of agent organizations.
Article
Cooperative Information Gathering Systems (CIGS) require the coordination of multiple agents executing tasks in a distributed, open and heterogeneous setting. In this complex and highly interactive context, it is essential to include explicit coordination models to ensure the coherent and efficient behaviour of cooperative agents. The paper identifies the three main issues which are essential for addressing this problem and expresses them in three interrelated models. The information model is implemented in Protege as a combination of an OWL ontology and SWRL rules. The task and organisational models are represented by High Level Petri Nets and extended And/Or graphs.